Exploring Reaction Centers in Retrosynthesis through Node Classification using Graph Neural Networks

The identification of the reaction center is a crucial step in the retrosynthesis of semi-template methods. Indeed, the existing methods often treat the task as a link prediction problem at the edge level. However, this approach may encounter challenges when dealing with cases involving multiple reaction centers. We approach the task from a node-level perspective, treating the identification of reaction centers as a node classification task. We employ advanced graph neural networks to capture molecular properties and topological structure. Then we apply FNN for node classification and adopt Focal Loss to address sample imbalance. Finally, through experimental validation, we demonstrate the feasibility of this method and achieve satisfactory results.